Beruflich Dokumente
Kultur Dokumente
= y
0
+ z
y
k
+ [x
+ e
iiJ
~
N(u, o
s
2
)
,
In the above, y
denotes the average growth rate of GDP of country i over a certain year range. In
line with Levine and Renelt (1992), z
is
a certain variable of interest potentially important explanatory variables of growth.
3
The cross-country growth regression model differs in an important way from models that use
panel data such as Savvides (1995) and Hoeffler (2002). These models that incorporate panel
data tend to address some issues that single cross-country regressions may have. Some of these
issues as pointed out in Hoeffler (2002) include the issue of reducing the time series to a single
(average) observation; omitted variable bias issue and endogeneity of some of the regressors.
Also, these models are used to capture country-specific effects. However, some of these issues
may not be as pronounced in the single cross-country regressions. For example, the bias of using
a single (average) observation may be small if the variable has not changed much over time as is
the case for some of the variables that are included in the economic growth literature.
4
Also,
endogeneity problem is usually addressed by using the initial values of the variables that may be
endogeneous in the model.
Attempt to solve the omitted variable bias has however led to an influx of variables that has been
included over time with the norm of looking at variables that are significant to determine the
factors that explain differences in growth rates across countries. This has led to the literature
addressing uncertainty in the variables to be included in these models. Levine and Renelt (1992),
Sala-i-Martin (1997a and b), and Fernandez, Ley, and Steel (2001) all investigated the issue of
model uncertainty. Fernandez, Ley, and Steel (2001) used a Bayesian framework that allowed
them to deal with both model and parameter uncertainty using Bayesian Model Averaging.
Ignoring the issue of using averages, the single cross-section growth regression specification
appropriately models differences in growth patterns of countries when there is no correlation
between the variable of interest and other explanatory variable. However, when the variable of
interest is potentially correlated with unobserved variables, the single cross-section growth
regression specification will lead to inconsistent estimate of the variable of interest. In the
following section, we describe a Bayesian estimation algorithm which properly accounts for the
impact of correlation between unobserved variables and the outcome of interest. This
specification is important for us to study the impact of climate change on economic growth.
)
-ypically, the estimation involves varying the pool of potentially important e.planatory variables of growth.
*
t can be argued that variables such as school enrolment, population growth and labor force has not
significantly diverged from the norm over a span of the sample period used in many of the growth studies.
$
3.2 Linear Hierarchical Model
Using Bayesian approach, this paper first assumes that climate change variables such as
temperature will have a different impact on GDP across countries and should be permitted to
vary across countries. There is however a degree of commonality across the continent on its
impact drought in South Africa will have an impact on the economy of neighboring countries
even if it was not as severe as that of South Africa. On the other hand, climate variables may also
have an impact on many of the explanatory variables that may be included (observed) or not
included (unobserved) in the regression equation. Consistent estimate of the parameters of
temperature and observed explanatory variables such as initial GDP per capita will require that
these variables be uncorrelated with the unobserved variables. This condition is unlikely to hold
especially given unavailability of data for many of the variables that can potentially influence
economic growth and related to temperature. This is the classic omitted variables bias and
inconsistency problem.
5
We propose a linear hierarchical model that is similar to the non-Bayesian fixed effects model
but exploits the hierarchical prior framework to estimate the parameters of the observed variables
that influence economic growth. The model proposed is in the spirit of the normal hierarchical
linear model described in Lindley and Smith (1972) and is similar to the model in
Abidoye, Herriges, and Tobias (2012) controlling for observed and unobserved variables using
country specific constants.
6
In particular, we will introduce a country-specific constant term that
captures both the observed explanatory variable and unobserved explanatory variable.
That is, we can employ the model:
y
t
= o
+ x
t
[
+ e
t
i = 1,2, , N; t = 1,2, , I
Where
o
= y
0
+ z
o
y
k
+ z
u
This problem resolves the omitted variable bias since e
t
is no longer correlated with the variable
of interest (x
t
). However, the impact of the observed explanatory variables on economic growth
will not be separately identified in the classic fixed effects specification.
We will estimate the above equation in a Bayesian framework and will adopt the blocking
strategy in Abidoye, Herriges, and Tobias (2012) by proceeding in a manner that is similar to the
+
Abidoye, /erriges, and -obias "2#12% illustrate this problem in a 0andom 1tility 2a.imization setting but
the setting is similar to ours by replacing choice alternatives with time.
,
'etailed description of this model and similar hierarchical models in the 3ayesian framewor4 can be found
in 5oop, !oirier, and -obias "2##$%.
6
classic fixed effects model by isolating the impact of the unobservable (capturing them entirely
in the country- specific constants) and insulate the climate parameter from their effects.
7
3.3 Hierarchical Priors
As stated in the previous section, the country-specific constants capture explanatory variables
that are included and not included in the regression that might explain the differences in
economic growth rates across countries.
8
The interactions of all country level variables that are
not of interest but typically included in cross-country growth models are solely captured in the
country-specific constants. We are also interested in estimating the correlation between the
climate variable and the unobserved variables that may not be captured in the regression. This
correlation will indicate the impact increase in temperature will have on these variables.
In our Bayesian approach, we capture the above by introducing a hierarchical structure into our
model, by assuming that each country shares some degree of commonality in their temperature
and economic growth by assuming that the country-specific constant and parameter on
temperature are drawn from the same normal population. That is, we allow for correlation
between the impact of temperature and other factors that may influence economic growth.
Specifically:
0
= j
o
[ ~ N(0
0
,) . (2)
Where:
0
0
= _
y
0
+z
o
y
k
[
0
_ = j
z
y
[
0
[
=_
_
uu
_
u[
_
u[
_
[[
_ = _
o
u
2
po
u
o
[
po
u
o
[
o
[
2
_
z
includes a constant term and the observed/included explanatory variables that influence
growth in country i. The correlation between temperature and the intercept is captured with p.
There are some silent features of our model that is worth mentioning our specification, as is the
case with most cross-country growth model will not solve the problem of potential correlation
between the included explanatory variables and the unobserved variables. It is typically assumed
$
As is pointed out in Abidoye, Herriges, and Tobias (2012), this simply echoes standard result that the fixed effects
estimator is unbiased even when correlation exists between the fixed effects and other explanatory variables
included in the model.
8
Also, the interactions of all country level variables that are of interest but typically included in cross-country
growth models are solely captured in the country-specific constants.
7
that this assumption holds. However, if this assumption does not hold, our specification can be
extended to make use of instrumental variables approach to consistently estimate y. In this paper,
we are particularly interested in consistently estimating [
and [
0
. Even when such correlation
exists, the inclusion of country-specific constants and our posterior simulator will yield
consistent estimates of the parameters of interest.
To complete our model, we specify priors for the remaining parameters. These are enumerated
below:
y~ N(p
y
, I
y
)
[
0
~ N(p
[
, I
[
)
-1
~ w(|p
0
R]
-1
, p
0
)
o
s
2
~ I0(o
s
, b
s
)
The hyper-parameters of the priors above, such asp
y
, I
y
, p
0
, o
s
, b
s
e.t.c., are supplied by the
researcher and are in general chosen to be relatively vague to allow dominance of the
information from the data. The notation N refers to the normal distribution, whereas W(.,.)
represents a Wishart distribution and IG(.,.) represents the inverse gamma distribution. There are
parameterized as in Koop, Poirier, and Tobias (pp. 336-339). These particular families of priors
are chosen primarily because when combined with the likelihood function yield conditional
posterior distributions that are easily recognized and sampled. These proper priors also make
model comparison and calculation of Bayes Factor relatively easy. Our prior means p
y
anu p
[
are set to zero vectors with the respective variance I
y
anu I
[
set to identity matrix and 25
respectively. The priors (hyperparameters) on the variance term are also selected by choosing
o
s
= S and b
s
= 1(4u).
9
p
0
is set to be equal to 5 and the prior is chosen to reflect some degree
of variability in the temperature and economic growth across countries. All these priors are
chosen to be reasonably diffuse and non-informative.
3.4 The Posterior Simulator
We fit the model using the Gibbs sampler and employ a number of blocking steps to mitigate
autocorrelations and consistently estimate our parameters of interest. Before describing these,
first let 1 = |{0
]
=1
n
y [
0
-1
o
s
2
] and define 1
-o
as all the elements of 1 other than .
The joint posterior distribution for all the parameters of this model can be written as:
7
-his chooses the prior mean for sigma82 e9ual to 2# with standard deviation also e9ual to 2#
1#
p(1|y)
__p(y
|H
, 0
, o
s
2
)p(0
|y, [
0
,
-1
)
N
=1
_ p(y|Z, p
y
, I
y
) p([
0
|X, p
[
, I
[
) p(o
s
2
|o
s
, b
s
) p(
-1
|p
0
, R)
Step 1: Draw {0
]
=1
n
|1
-{0
i
]
, y
This complete conditional is proportional to the joint posterior distribution p(1|y). Absorbing all
the terms that do not involve 0
= _
y
1
y
2
.
y
t
_, H
= _
1 x
1
1 x
2
. .
1 x
1
_.
Thus we obtain:
p(0
|1
-0
i
, y)~N(
0
i
J
0
i
,
0
i
), i = 1,2, , N
Where
0
i
= _
H
i
H
o
s
2
+
-1
_
-1
J
0
i
=
H
i
y
o
s
2
+
-1
0
0
We sample each of the 0
|y, [
0
,
-1
)
N
=1
_ p(y|Z, p
y
, I
y
)
Once we condition on the 0
Thus we can write:
y|1
-y
; y~N(
y
J
y
,
y
)
Where
y
= (
z
i
z
vai(u)
+ I
y
)
-1
And
J
y
=
z
i
o
vai(u)
+ I
y
p
y
Step 3: Complete Posterior Conditional for [
0
The complete Posterior Conditional for [
0
is similar to that of y above. Once we condition on
the 0
s ([
[
0
= (
N
vai(:)
+ I
[
0
)
-1
And
J
[
0
=
_ [
Nvai(:)
+ I
[
0
p
[
0
12
Step 4: Complete Posterior Conditional for o
s
2
o
s
2
|1
-c
s
2; y ~I0 _N -
I
2
+ o
s
, ju.S (y
- H
)
i
(y
- H
) + b
s
[
-1
]
Step 5: Complete Posterior Conditional for
-1
-1
|1
-
-1; y ~ w _j(0
- 0
0
) (0
- 0
0
)' + Rp
0
[
-1
, N + p
0
]
4. Data, estimation techniques, descriptive statistics and analysis of results
4.1 The Data
This section describes the data used to run the models specified above and the descriptive
analysis presented in the next section. Temperature data for each African country was gotten
through the Climate Research Unit (CRU). The study used observed gridded monthly mean
temperature data from the Climate Research Unit (CRU, version 3.0, Mitchell and Jones 2005).
The CRU dataset is based on station data and has a 0.5X0.5 resolution. The Global Gridded
Climatology data is presented at a new high resolution and made available by the Climate
Impacts LINK project, Climate Research Unit, University of East Anglia, Norwich, UK
(Mitchell and Jones, 2005). The Climatic Research Unit (CRU) data set is composed of monthly
0.50 latitude/longitude gridded series of climatic parameters over the period 1901-2009 however
the data used for this paper runs from 1961-2009.
For the purpose of studying the impact of climate change on economic growth in Africa, we find
it suitable to use data from the Africa Development Indicators (ADI) (2011) publication of The
World Bank. Economic growth is measured as the Annual percentage growth rate of GDP at
market prices based on constant local currency. Population data was also obtained from ADI.
Total population is based on the de facto definition of population, which counts all residents
regardless of legal status or citizenship--except for refugees not permanently settled in the
country of asylum, who are generally considered part of the population of their country of origin.
The values shown are midyear estimates.
Human capital investment is proxied for by primary school enrolment rates and life expectancy.
Although previous research (e.g. Mankiw, Romer, and Weil (1992) and Gemmell (1996)) has
argued that using the level of human capital with school enrolment can be problematic, we still
include it in the estimation. It has been used in many other studies and we let the model
likelihood dictate if it should be included or not.
1)
The Data is available for 34 countries. The sample consists of: Algeria, Benin, Botswana,
Burkina Faso, Burundi, Cameroon, Central African Republic, Chad, Congo, Dem. Rep., Congo,
Rep., Cote d'Ivoire, Egypt, Arab Rep., Gabon, Ghana, Kenya, Lesotho, Liberia, Madagascar,
Malawi, Morocco, Niger, Nigeria, Rwanda, Senegal, Sierra Leone, South Africa, Sudan,
Swaziland, Tanzania, Togo, Tunisia, Uganda, Zambia, and Zimbabwe.
4.3 Estimation and Testing
The algorithm described in Section 3 has been used to run our posterior simulator for 100 000
iterations discarding the first 5 000 of these as the burn-in. Results from these runs suggested that
the chain mixed reasonably well and appeared to converge within a few hundred iterations.
Although our point estimates are suggestive of good performance, any Makov Chan- Monte
Carlo (MCMC)-based inference can be affected by the degree of correlation among the
parameter draws over sequential iterations. The mixing of the posterior simulations has been
used to determine how many draws are needed to achieve the same level of numerical precision
that would be obtained under an independent identically distributed (iid) sampling. A high
degree of correlation will lead to a slow mixing that may not let us explore all areas of the
posterior as needed. These inefficiency factors, as they are called can be calculated by using the
definition of the numerical standard errors (NSE) of a Monte Carlo estimate with correlated
draws. The mean estimates can be obtained as:
NSE(0
m
) =
_
o
2
m
_1 + 2 _1 -
]
m
] p
]
,
m-1
]=1
Where 0 represents an arbitrary scalar parameter of interest, m denotes the number of post-
convergence simulations, 0
m
represents our estimate of E(0|y) as the sample average of our
post-convergence draws, p
]
represents the correlation between simulations ] periods (iterations)
apart and o
2
Io(0|y).
The NSEs for our models are extremely small relative to the mean estimates which strongly
indicate our simulation based estimates accurately approximate the posterior means of this
selection parameters. This, again, suggests that our algorithm mixes quite well.
10
1#
:e do not present the values for the ;&<s but all of them are less than #.##+.
1*
4.3 Descriptive Analysis
This section presents the main feature of temperature dynamics in the 34 African countries used
in this paper. -able 1 presents the minimum and maximum temperatures, the difference between
the minimum and maximum, the mean (1961 and 2009) and the absolute change between 1961
and 2009. Based on the mean value, Burkina Faso, Senegal, Benin, Niger and Ghana are among
the hottest countries in Africa while Lesotho, Morocco, South Africa, Rwanda and Tunisia
appear to be the coldest. Sudan, Botswana and Niger experienced the highest swings between
the minimum and the maximum temperature over the period of 49 years. Countries that changed
by more than 2
o
Celsius between 1961 and 2009 are Sudan (3.04), Chad (2.61), Niger (2.47) and
Egypt (2.15).
Figure 1 shows the trend of temperature for countries with the highest swings over the period.
Sudan and Chad have the highest levels and have been rising consistently between 1961 and
2009. They are followed by Uganda, Botswana and Tunisia. Countries that experienced some
relative stability in temperature between 1961 and 2009 include Madagascar, Congo Democratic
Republic, Gabon, Liberia and Sierra Leone (see Figure 2).
As shown in Figure ), lag of temperature change appears to have inverse relationship with the
change in current output. This is a clear indication that lag of change in temperature is a good
predictor of change in the level of outputs. A similar trend is observed for agriculture (Figure *).
The pattern for most countries follows the regional trend as shown for Sudan in Figure +. The
correlation index between temperature and agriculture value added is -0.61.
Table 1: Descriptive analysis of temperature (1961-2009)
Row Labels Min Max Max -
Min
Mean
Temperature
Standard
errors
Absolute
change in
temperature
(1961 2009)
Algeria 21.72 24.04 2.32 22.96 0.55 1.01
Benin 26.62 28.61 1.99 27.56 0.46 1.02
Botswana 20.39 23.21 2.82 21.86 0.62 1.46
Burkina Faso 27.54 29.12 1.58 28.32 0.39 1.34
Burundi 19.83 21.73 1.91 20.48 0.46 0.96
Cameroon 24.00 25.51 1.51 24.71 0.33 1.01
Central African
Republic
24.28 26.02 1.74 25.10 0.45 1.06
Chad 25.72 28.33 2.61 26.99 0.58 2.61
Congo, Dem. Rep. 23.79 25.33 1.54 24.62 0.30 0.64
Congo, Rep. 23.75 25.10 1.35 24.23 0.33 1.01
Cote d'Ivoire 25.58 27.17 1.59 26.41 0.32 0.21
Egypt, Arab Rep. 21.54 23.74 2.19 22.57 0.56 2.15
1+
Gabon 24.17 25.91 1.75 25.09 0.31 0.46
Ghana 26.45 28.14 1.70 27.29 0.37 0.68
Kenya 23.49 25.55 2.06 24.59 0.43 1.06
Lesotho 11.48 13.40 1.92 12.39 0.49 0.49
Liberia 24.71 26.10 1.39 25.38 0.29 0.42
Madagascar 21.67 22.81 1.14 22.30 0.32 0.05
Malawi 21.20 22.91 1.71 22.01 0.40 0.71
Morocco 16.04 18.47 2.43 17.36 0.53 0.29
Niger 26.20 28.68 2.47 27.45 0.49 2.47
Nigeria 26.19 27.84 1.65 26.93 0.38 1.52
Rwanda 18.32 20.24 1.92 18.99 0.48 1.09
Senegal 27.14 29.06 1.92 28.08 0.46 0.47
Sierra Leone 25.60 26.97 1.37 26.25 0.32 0.60
South Africa 16.96 18.60 1.64 17.85 0.42 0.82
Sudan 25.82 28.86 3.04 27.26 0.73 3.04
Swaziland 19.47 21.16 1.68 20.21 0.44 0.34
Tanzania 21.83 23.38 1.55 22.52 0.42 0.66
Togo 26.24 28.27 2.04 27.19 0.44 0.84
Tunisia 18.40 20.87 2.47 19.71 0.68 1.14
Uganda 22.01 24.58 2.57 23.00 0.67 1.90
Zambia 20.96 23.29 2.33 21.84 0.52 0.92
Zimbabwe 20.29 22.91 2.62 21.28 0.56 1.14
4.4 Analysis model of results
The analysis of the link between temperature and economic growth is based on the common
intercept
0
, common slope
0
, variance parameters of the second-stage covariance (denoted
by
2
and
2